CN108351432A - The focusing inverting of data-driven - Google Patents

Abstract

The inverting of the focusing of data-driven discloses a kind of method for the geophysical data for handling and being acquired from subterranean zone.This method includes receiving the geophysical data acquired from subterranean zone, determine the data point high to quite sensitive degree, point executes inverting based on the data, wherein, during inverting, it is realized to abnormal focusing by increasing the weight of the data point, and determines at least one model that physical attribute is distributed in subterranean zone.

Description

Data-Driven Focused Inversion

technical field

The present invention relates to geophysical surveying and data processing, and more particularly to methods for optimizing physical property distribution models based on data acquired from geophysical surveying methods such as controlled source electromagnetic or seismic surveying.

Background technique

Geophysical data collected from subsurface regions via conventional acquisition methods (e.g., surveys such as seismic or CSEM) can be used to obtain information about the physical properties distribution information. Using the acquired data and prior knowledge about the subsurface region, a model characterizing the distribution of physical properties can be generated. In this term, "model" refers to an estimated quantity. Inversion involves the optimization of the resulting model in order to determine the distribution of properties that explain the observed data.

The inversion problem (ie, inversion of model parameters from a measured data set) is almost always an undetermined, non-unique and nonlinear problem (Tarantola, A. Inverse Problem Theory, Elsevier, New York 1987; Tarantola, A. Inverse Problem Theory and Methods for Model Parameter Estimation, Society for Industrial and Applied Mathematics 2005, and Zhdanov, M.S. Geophysical Inverse Theory and Regularization Problems, Elsevier 2002). The measured data does not provide enough information to ensure a unique solution at the desired resolution. In addition, measurement data is always subject to noise, inaccuracies and errors. Finally, the implementation of the forward modeling operators required by the inversion engine often has limited accuracy because they are based on assumptions for practical and computational reasons. To obtain a solution, artificial constraints on the model parameters and regularization are usually introduced to stabilize the inversion problem (Tarantola, 2005 and Zhdanov, 2002). The assumptions behind these constraints and regularization are often imprecise.

Therefore, the resulting final inversion model is a compromise between all these limitations, imprecision and errors. In practice, the inversion process always ends up in a region of attraction with corresponding local minima that favors certain parts and aspects of the inversion model at the expense of others. Which field of attraction and corresponding local minima the inversion run actually ends up with depends on the initial model, the regularization parameters, the data points included, the weights of the individual data points, and the optimization algorithm used. The outcome of this struggle between the different inversion drivers is unmanageable without adding very strong assumptions/constraints, potentially leading to severe artifacts in the inversion results if the assumptions are imprecise.

It is often desirable to reconstruct certain parts and aspects of a model as accurately as possible. The cost of less accurate reconstruction of other less important parts/aspects of the model can then be accepted. Typically, the size of the survey target is very small compared to the entire model to be inverted. Correspondingly, the number of data points that are mainly sensitive to the large model background is completely dominated by the number of data points that are actually sensitive to the target. Consequently, the target response is often too weak to compete with other, much stronger inversion drivers, resulting in poor imaging of the target and reasonably good imaging of the background.

As mentioned above, constraints and regularization are routinely used to focus the inversion search space and results (Tarantola and Zhdanov 2002), e.g., encouraging the inversion engine to facilitate certain aspects of the inversion model, e.g., smoothness, density anomalies , low/high values for some parts of the model (Tarantola and Zhdanov 2002). In the following, we use the term regularization to refer to both constraints and regularization. This conventional approach is model-driven focused inversion (MDFI): it is applied directly to the properties of the model. The problem with this type of model-driven regularization is that the inversion model is forced to contain the properties required for regularization, whether or not those properties are present in the real model.

A model-driven focused inversion (MDFI) method can be shown as follows. Let the inversion/optimization problem find the M-dimensional model vector that minimizes the following mismatch function

Here, the data residual in, is a real or complex N-dimensional vector representing the measurement data, while is for the model forward simulation data. reverse vector notation Represents the conjugate transpose. Data covariance matrix Describe data errors and their correlations. Conventionally, by a diagonal matrix containing the variance of the data to Make an approximation. diagonal Indicates that all data points are irrelevant. The routine used in geophysical inversion Examples can be found e.g. in (Morten, JP, AK, T. (2009) CSEM data uncertainty analysis for 3D inversion, SEG Expanded Abstracts 28, 724).

Regularized mismatch function of Depends only on model parameters. λ is a real scalar used to control the strength of the model regularization term. Typically, technicians use to stabilize the inversion problem and control the inversion flow, i.e. enforce/encourage the inversion model to have certain properties such as smoothness, sharpness, parameters kept within bounds, proximity to a certain priority model, etc. (Tarantola, 2005 and Zhdanov, 2002).

Contents of the invention

A Data-Driven Focused Inversion (DDFI) method is proposed, which encourages the inversion optimizer to focus more on some specific attributes, parameters, and regions of the inversion model while sacrificing other attributes, parameters, and regions. Unlike MDFI, DDFI does not force the inversion model to include the focused attribute if it does not exist in the data.

Description of drawings

Figure 1 shows the vertical (R _V ) and horizontal (R _H ) resistivity anomalies for the first and second forward models.

Figure 2 shows the electric field amplitude (E _x , solid line), phase (dots), and noise estimates for the azimuth receiver.

Figure 3 shows the focusing matrix of the axial component _Ex of a drag chain with index n=1/2 and f=1Hz of the diagonal.

Fig. 4 shows the real model T, the initial model C, the vertical resistivity model inverted by α=0, and the vertical resistivity model inverted by α=20.

Fig. 5 shows the real model T, the initial model C, the horizontal resistivity model inverted by α=0, and the horizontal resistivity model inverted by α=20.

Fig. 6 shows the inverted vertical resistivity model with α=0 and α=20 inversion for the real model B, the initial model C.

Figure 7 shows the inverted vertical resistivity model for the initial model C, inverted with α=0, α=5 and α=20.

Figure 8 shows the vertical resistivity model inverted with α=0 and α=5 for the true model T, the initial model D.

Figure 9 shows the vertical resistivity model inverted with α=0 and α=20 for the true model T, the initial model E.

Figure 10 shows a flowchart of an exemplary method.

Fig. 11 schematically shows an exemplary computer device in block diagram form.

Figure 12 is a schematic diagram of a computer system for implementing one or more embodiments of the invention.

Detailed ways

The method according to the present invention implements geophysical data processing techniques to generate a physical property distribution model of a subsurface region. Geophysical data acquired from a geophysical survey and prior knowledge about the subsurface region are used to generate a model characterizing the distribution of physical properties that produced the acquired data.

Some data points are highly sensitive to certain parameters, attribute regions/parts of the model. By increasing the weight of these data points during the inversion process, a more accurate model of the distribution of physical properties can be achieved, e.g. properties not present in the data will not be part of the inversion model.

Inversion involves adding new terms to the covariance matrix to encourage the inversion engine to focus its efforts on minimizing mismatches in selected data residuals that are highly sensitive to certain parameters, properties, regions/parts of the model. Hereinafter, we will use the term target to denote a parameter, property or region/part of the desired model that we wish to study. We introduce a modified covariance matrix, which is defined as:

here, is the modified covariance matrix, is the identity matrix, is the focusing matrix, and α is a constant used to control the focusing strength. In order to achieve focus on the goal, should be defined as small/zero for data points that are not sensitive to the target, and large for data points that are sensitive to the target. In order to achieve focus on the target, it is assumed (given) in the matrix In , the most sensitive data points should have a higher weight (higher focus) compared to the data points that are not sensitive to the target.

In another embodiment of the invention, the focusing matrix can also be used in the reverse sense, i.e., the Defined as being small for data points sensitive to certain parameters, properties, regions of the inversion model and large at the remaining data points. This encourages the inversion engine to focus its least effort on non-target parts. This inverse approach can be used to reduce the influence of extraneous parts of the model in the inversion. Depending on the purpose of the inversion study, "irrelevant" parts of the model could be, for example, salt, basalt, pipelines, etc.

one for building The method is through the scattered field. Here two forward simulations are performed: one for the background model which has simulated data once for model It contains the same background model with added targets. Then the scattered field is defined as:

therefore, Respond for data only from the target. Then, can for example be defined as:

is a user-supplied matrix function, e.g.,

where n is the exponent provided by the user. Matrix functions for user-supplied scattered fields, for example:

in, express The outer product, and the matrix function diag() extracts the diagonal matrix.

To be careful of, The form ensures defocusing of noisy data points.

for diagonal and Equation 7 acts much like adding weight to the data points where the scattered field is strongest.

By way of example, a synthetic data set can be used to illustrate the effectiveness of the present invention and show that a data-driven focused inversion approach improves inversion results around targets; this approach is applicable to synthetic seafloor logging controlled source electromagnetic (CSEM) Inversion of the dataset.

Electromagnetic data consists of electric and magnetic fields: [E _x ,E _y ,E _z ] and [H _x ,H _y ,H _z ], respectively. The components in the electromagnetic field vector are complex in the frequency domain and depend on the source (s), receiver (r) and frequency (f), eg, _Ex → _Ex (s,r,f).

Conventionally, the data covariance matrix used in CSEM inversion is assumed to be diagonal with elements given by the data uncertainty factor, examples can be found in (Morten, 2009; FA and Nguyen, AKE Enhanced subsurface response for marine CSEM surveying, Geophysics, 75 No. 3, pp. 7-10, 2010). Here we define the variance for different electromagnetic fields as:

where _β1 is the autocorrelation multiplicative uncertainty between the field components, and _β2 and _β3 are the cross-correlation multiplicative uncertainties between the field components. Furthermore, η is the field-additive noise alone. Maao and Nguyen (2010) provide the reasons and explanations behind multiplicative uncertainty and additive noise. Thus, a single i-(source, receiver, frequency) measurement is a data vector Up to 6 additional components are provided, and then the corresponding diagonal covariance matrix for the electric field part is:

Similarly for the magnetic field is:

Synthetic Data Example

In this section we show that the data-driven focused inversion method is well suited for 3D anisotropic inversion on synthetic seabed log controlled source electromagnetic datasets.

Synthetic models and synthetic datasets are largely generated from measurement data from real surveys. The synthetic model is created using the real seismic horizons at the surveyed location. Seismic-driven CSEM inversion from real datasets (Nguyen, AK, Hansen JO (2013) CSEM exploration in the Barents Sea, Part II—High resolution CSEM inversion, 75th Annual Conference and Exhibition, EAGE, Extended Abstracts, We0403) to extract the resistivity value. mistake! (Reference source not found) Slices of the ground truth model (T) and the background model (B) are shown. Model T has vertical resistivity R _V anomaly in it; while there is no anomaly in Model B. This _RV anomaly in the real model is the target of an inversion study. The horizontal resistivity model R _H is the same for both Model T and Model B. The reason behind the lack of anomalies in R _H stems from the fact that the thin horizontal resistance anomalies have practically no effect on the CSEM data, so no R _H anomalies were constructed in the earthquake-driven CSEM inversion. We do not think it is necessary to artificially add R _H anomalies to the real model. The 3D forward modeling of the grid dx=dy=100m and dz=40m is based on the finite difference time domain method ( FA (2007) Fast finite-difference time-domain modeling for marine-subsurface electromagnetic problems, Geophysics, 72, A19-23). Data were collected for frequencies f = 0.25 Hz, 0.5 Hz, 1 Hz. Bathymetry was considered in the modeling.

Transmit and receiver navigation and orientation are derived from survey data. For each synthetic data point, we add to it a complex number with random phase and amplitude given by the noise estimate for the corresponding real data point. Furthermore, we distort the receiver orientation values by uniformly distributed random numbers between -3 and 3 degrees. Therefore, noise levels, receiver rotation errors, and navigation imperfections are all close to real acquisition uncertainties and conditions. The noise-influenced electric field of the azimuth line is wrong! (Reference source not found). We see that additive noise can significantly affect the data in both amplitude and phase, especially at offsets where the magnetotelluric burst occurs.

In order to construct the target-focused covariance matrix, the sensitivity of each data point to the target needs to be estimated. This can be achieved by comparing clean synthetic data from model T and model B.

mistake! (Reference source not found) shows the focusing matrix of index n=1/2 and f=1Hz of the axial component E _x of a drag chain of the diagonal. We see that targets clearly project their presence and information into a limited set of data points.

inversion

L-BFGS-B scheme (Zach, JJ, AK, T., and F. 3D inversion of marine CSEM data using a fast finite-difference time-domain forward code and approximate hessian-based optimization. SEG Expanded Abstracts 27, 614, 2008) for inversion on synthetic datasets from Model T and Model B. No regularization was applied in this pixel-based focus inversion. The initial model C is a smoothed and enlarged version of model B with dx = dy = 150m and dz = 40m. For the inversion, we use fields E _x and E _y and frequencies f = 0.25 Hz, 0.5 Hz, 1 Hz, and the corresponding maximum absolute offsets are 10000 m, 9000 m, 8000 m. For all frequencies, the minimum offset is 1500m. Unless explicitly mentioned otherwise, the final inversion results are obtained by iteration 100.

mistake! (Reference source not found) shows the real, initial model C, vertical resistivity model inverted with α=0, and vertical resistivity model inverted with α=20. We see that the model inverted by α=0 contains anomalies at the correct locations. However, the reconstructed anomalies are rather weak at very low resistivities. Furthermore, we see that the model for α = 20 inversion reconstructs the anomaly much better than the conventional inversion. The difference in resolution between the real model and the inversion model stems from the fact that the real model is based on seismic-guided inversion, whereas the inversion model here is pure CSEM data at a much lower resolution than in the seismic-guided case Driven pixel-based inversion.

mistake! (Reference source not found) shows the real, initial model C, the horizontal resistivity model inverted by α=0, and the horizontal resistivity model inverted by α=20. We see that the inversion model is more or less the same and very similar to the initial model. Even the receiver imprint is more or less the same. Most likely, these receiver signatures are the signature of introduced errors in the direction of the receiver. Note again that the introduction of thin, high-resistance anomalies in the true level model does not affect the data and thus the inversion results.

The natural question to ask is whether focusing on anomalous weights might introduce artifactual anomalies in the inversion results, even when there are no such anomalies in the input data. Figure 6 shows the inversion model for α=0 and α=20 from dataset B. Here, the real model has no resistivity anomaly. We see that all inversions correctly predict no resistivity anomalies in the model. Therefore, it can be concluded that weighting of focused anomalies does not artificially introduce resistivity anomalies that are not present in the input data.

Figure 7 shows the inverted vertical resistivity model for α=0, α=5 and α=20. It can be seen that increasing the focus also improves the clarity of the resistivity anomalies in the inversion model.

Another natural question is how does a significantly wrong initial model affect the inversion results when applying weights that focus on anomalies? We investigate this problem by performing an inversion using initial models D and E, which are distorted versions of the "correct" initial model C. The initial model D is a modified version of model C with R _V → R _V × 0.8 and R _H → R _H × 0.8.

mistake! (Reference source not found) Vertical resistivity models are shown for: real, original model D, inverted with α=0, and inverted with α=5. We see that when focused weights are used in the inversion, the resistivity anomaly is again recovered much better compared to conventional unfocused weights. The inverted horizontal resistivity model corresponding to the model shown in Fig. 8 is more or less the same.

mistake! (Reference source not found) 9 shows the vertical resistivity models for: real, original model E, inverted with α=0, and inverted with α=5. The initial model E is a modified version of model C with R _v → R _v × 1.2 and R _H → R _H × 1.2. We see that when focused weights are used in the inversion, the resistivity anomaly is again recovered much better compared to conventional unfocused weights. The inverted horizontal resistivity model corresponding to the model shown in Fig. 8 is more or less the same.

Therefore, it can be concluded that a significantly wrong initial model does not reduce the effect of focusing on the weight of anomalies.

Above we have shown that the Data Driven Focused Inversion (DDFI) method works well for CSEM inversion.

Unlike conventional model-driven focused inversion (MDFI) methods, DDFI does not impose certain properties into the inversion model if their presence is not present in the data.

Inversion of measured data often provides results with lower resolution than required for interpretation, or estimates parameters with higher uncertainty than required, especially for small target areas or small groups of important parameter. Data-driven focused inversion puts more effort into reducing data residuals at selected data points that are sensitive to user-defined objectives. Recall that the term target here refers to selected properties, parameters, and regions of the model. Thus, the inversion result will provide better sharpness of the target at the expense of a poorer background which is usually acceptable from the user's point of view.

Inversion almost always ends with a local minimum. Which minimum it ends up with depends on the initial model, regularization, optimizer algorithm. Model-driven focused inversion imposes on the inversion model the properties indicated by the regularization term, e.g. the smoothness regularization term forces the inversion model to be smooth regardless of the measurement data. In contrast, data-driven focused inversion only puts more effort into reducing data residuals at selected data points. Therefore, the focusing function for the target does not force the inversion model to include the target if no target response is found in the measurement data. Instead, DDFI in this case will ensure that the target is not included in the inversion model because the weighted data mismatch would be too high. Therefore, DDFI increases the probability that the inversion ends with a local minimum, which ensures that the properties of the target are accurately described.

Data-driven focused inversion DDFI attempts to find a model that minimizes the data fit for selected data points that are sensitive to the target at the expense of other data points that are only sensitive to the background and not to the target. This enables more precise definition of the target.

Fig. 10 is a flowchart showing exemplary steps according to the first embodiment. The following numbering corresponds to that of Figure 10:

S1. Receiving geophysical data collected from a subterranean region.

S2. Identify data points with high sensitivity to anomalies.

S3. Performing an inversion based on the data points, wherein focusing on anomalies is achieved by increasing the weight of the data points during inversion.

S4. Determining at least one model of distribution of physical properties in the subterranean region.

Figure 11 is a flowchart showing exemplary steps according to the second embodiment; the following numbering corresponds to that of Figure 11:

S5. Receiving geophysical data collected from the subterranean region.

S6. Identify data points with low sensitivity to anomalies.

S7. Performing an inversion based on the data points, wherein during inversion focusing on anomalies is achieved by reducing the weight of the data points.

S8. Determining at least one model of distribution of physical properties in the subterranean region.

Referring to Figure 12, there is provided an exemplary computer device 1 capable of performing a method of processing geophysical data collected from a subterranean region, as described in embodiments of the present invention. The computer device may contain or be programmed by a computer program 2 comprising computer readable instructions for processing geophysical data collected from subterranean regions, as described in embodiments of the present invention. The computer program 2 may be stored on a non-transitory computer readable medium in the form of a memory 3 . The computer program 2 can be provided from an external source 4 . The computer device 1 may comprise a display 5 which allows a user to visualize a model of the distribution of physical properties in the determined subterranean region.

Claims (14)

1. A method of processing geophysical data collected from a subsurface region, the method comprising: Receive geophysical data collected from subsurface areas; Identify data points with high sensitivity to anomalies; performing an inversion based on the data points, wherein during inversion focusing on the anomalies is achieved by weighting the data points, and At least one model of distribution of physical properties in the subterranean region is determined. 2. The method of claim 1, wherein the geophysical data may be any one of electromagnetic data, controlled source electromagnetic data, seismic data, gravity data, or resistivity data. 3. The method according to claim 1 or 2, wherein increasing said weighting of data points is achieved by defining a covariance matrix of the form: in, is the focusing matrix, and α is a constant used to control the focusing strength, and increasing the 4. A method according to any one of claims 1, 2 or 3, wherein a high sensitivity of a data point to an anomaly is determined by comparing two forward models except All the same except for the exception. 5. A method according to any preceding claim 4, wherein the anomaly may be any of a parameter, property or region/part of the model. 6. A method of processing geophysical data collected from a subsurface region, the method comprising: receiving geophysical data collected from the subterranean region; Identify data points with low sensitivity to anomalies; performing inversion based on the data points, wherein during inversion focusing on the anomalies is achieved by reducing the weight of the data points; and At least one model of distribution of physical properties in the subterranean region is determined. 7. The method of claim 6, wherein the geophysical data may be any one of electromagnetic data, controlled source electromagnetic data, seismic data, gravity data, or resistivity data. 8. A method according to claim 6 or 7, wherein increasing said weighting of data points is achieved by defining a covariance matrix of the form: in, is the focusing matrix, and α is a constant used to control the focusing strength, and increasing the 9. A method according to any one of claims 6, 7 or 8, wherein the low sensitivity of a data point to an anomaly is determined by comparing two forward models except All the same except for the exception. 10. A method according to any one of claims 6 to 9, wherein the anomaly may be any one of a parameter, property or region/part of the model. 11. A computer program arranged to control a computer to carry out the processing in the method according to any one of claims 1 to 10. 12. A computer comprising or programmed with the program according to claim 11. 13. A computer-readable medium containing the program according to claim 12. 14. Transmission of the program according to claim 11 via a network.